On-device · data firewall for AI

Catch the secret before it reaches the AI.

You paste a stack trace, a config file, a client's email into ChatGPT — and an API key or an IBAN rides along. Proxia inspects every prompt on your own device, masks what's sensitive, and restores it in the reply. Nothing is ever sent to a server.

Watch — it's already inspecting the panel. Type your own, or load an example.

chatgpt.com
try:
Checkpoint
The part that matters

Most "AI DLP" tools have to see your text. Proxia never does.

To check a prompt, a cloud filter has to receive it first — so your secrets pass through someone else's server on the way to being "protected." Proxia's engine runs inside your browser. The comparison is the whole product:

Cloud AI-DLP

Your text leaves to be checked.

You──▶their server──▶The AI

A second copy of every prompt now lives somewhere you don't control. The tool meant to protect your data is itself a place it can leak.

Proxia

Your text is checked in place.

You──▶Proxia · on-device──▶The AI

A pure Rust core compiled to WebAssembly inspects the prompt locally. No telemetry, no beacons. The only thing that ever leaves is what you decide to send.

What it catches

Trained on the things you'd regret pasting.

Regex plus algorithmic validation — Luhn, IBAN mod-97, national-ID checksums — keeps false positives low. A random 16-digit number isn't a card unless it checks out.

Secrets // critical

API keys and tokens, private-key blocks, passwords in config, connection strings.

  • sk-…, AKIA…, ghp_…
  • -----BEGIN PRIVATE KEY-----
  • PASSWORD=…
  • high-entropy strings

Financial // high

Bank details validated, not just pattern-matched.

  • IBAN (mod-97)
  • Card number (Luhn)

Identity // personal

National IDs with real checksums, plus contact data.

  • French SSN (NIR)
  • Belgian national number
  • International phone
  • Email

Manipulation // injection

Not just your data leaving — attacks coming in.

  • Prompt-injection phrasing
  • Hidden Unicode / bidi
  • Phishing signals
100%recall
0%false positives

Measured by a reproducible benchmark on a labelled corpus (64 positive / 89 benign cases), per detector, gated in CI. It's a coverage guarantee on the tested cases — not a promise of catching everything in the wild. Detection is strongest in French & English today.

The checkpoint

Three moments, one decision.

01 — write

You compose as usual

Type or paste into ChatGPT, Claude, Gemini — whatever you already use. Proxia stays out of the way.

02 — check

Inspected the instant you send

In ~40 ms, on-device, before a single byte leaves. If it's clean, you never notice.

03 — decide

Send masked, send anyway, or cancel

One click swaps secrets for tokens and restores them in the reply. Your call, every time.

No telemetry

No analytics, no beacons, no phone-home. Your prompts are never transmitted anywhere.

Your log, encrypted

An optional local audit records counts only — never the sensitive text — sealed with a key only you hold.

Non-destructive

Proxia flags, masks, and proposes. It never silently deletes your text — you keep the last word.

Beyond the prompt

It guards the tools your AI calls, too.

Agents don't just chat — they run shell commands, send email, hit APIs. A secret leaks through a tool call's arguments. Proxia inspects those and blocks a credential heading into a dangerous tool — on the browser, and directly on the MCP tool-call path.

Where it works

The AIs you already use.

ChatGPTClaudeGemini CopilotPerplexityMistral PoeDeepSeekGrok Meta AI+ MCP / tool-use
Where it runs

One engine, wherever the data shows up.

The same on-device engine ships as a browser extension and as Office add-ins. Nothing you inspect ever leaves the machine — whichever surface catches it.

Browser extension

Inspects the prompts you send to an AI, and masks sensitive values displayed on any page you add.

v1.0.0 Chrome · Edge · Safari

Excel add-in

Masks the sensitive cells — IBANs, cards, secrets — before you show the sheet in a call, a demo or an open office. Scan a selection, mask it, reveal a cell when you need it. Works in Excel desktop and Excel web, where a browser extension sees nothing at all. It protects the screen, not the file you send.

beta sideload · desktop & web

Outlook add-in

Checks the mail before it leaves your outbox, and warns you on the way out if a secret or a client's data rode along.

beta sideload · Mac & web

The Office add-ins are in private beta. Want to hear more, or try Proxia on your own spreadsheets and mailbox? Tell us what you'd put in front of it — we're looking for testers.

Get in touch →
Straight answers

The questions you're already asking.

Does it slow me down?
No. The check runs in about 40 milliseconds, on-device, at the moment you send. If your prompt is clean you never see anything.
What about false positives?
Detectors that can be validated are validated — a card must pass Luhn, an IBAN must pass mod-97, a national number must match its checksum. And with per-site profiles you can quiet a category on a specific site. Critical secrets are always flagged regardless.
Is my data really safe?
The detection engine is a pure Rust core compiled to WebAssembly that runs inside your browser. There is no network call in the inspection path — no prompt, ever, is sent to a server. The optional local audit log stores counts only, encrypted with a key only you hold.
Which languages does it cover?
Detection is strongest in French and English today; PII like IBANs, cards and emails is language-agnostic. Broader NL/DE coverage is on the roadmap — and the honest coverage limits are documented, not hidden.
What can't it see?
The browser extension only sees what the browser sends. For native desktop apps and CLIs, a local proxy covers the gap; for AI agents, the MCP guard sits directly in the tool-call path.

It's local. It's yours.

Paste something you'd normally hesitate to. Proxia is the safety net that lets you use AI without second-guessing every prompt.

$ ./extension/build.sh  # then load unpacked

Chrome Web Store · v1.0.0 · submission-ready